Skip to content
JacopKanePublic

About

A Claude Code plugin that puts the parts of your setup in one settings file, each with its own switch. Early work in progress.

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

hub

CI License: MIT

Every part of your Claude Code setup gets its own switch. A gate stops private text before it goes public.

hub is a plugin for Claude Code. It puts the parts of a Claude Code setup in one settings file and gives each part its own switch. There is no extra command to learn. You tell Claude what to change in plain words, such as "turn off the status band". The parts include the public text gate, agents for different kinds of work, writing rules, a status band above the prompt and a coordinator chat. The coordinator is one chat that passes your request to the chats that own each topic and reports back in one place.

The public text gate stops a commit message that holds a home path. Claude rewrites it and commits. Then Claude tries to switch the gate off, and hub refuses. Only you can lower a protection.

hub is an early work in progress and is not released yet. Expect changes that break things.

What works today and what is planned

  • Working today are the settings file and its switches, a first-run setup, self-protection, the public text gate, the status band, agent tiers, writing rules, and the coordinator chat with a chat board.
  • Cursor support is in beta and untested. It works only from a checkout of this repo, not from the plugin install.
  • Still planned are a sandbox command and a first release.

Install

You need Claude Code and Python 3, on macOS or Linux. hub is tested with Claude Code 2.1.319. On other versions, hub's setup check shows a warning, and the message queue sends each message at once instead of holding it.

claude plugin marketplace add JacopKane/hub
claude plugin install hub@hub

Then start a new chat. hub offers its setup there. A small Haiku agent reads your setup, asks a few questions in one dialog, and turns on a small starting set. Ask Claude "what does hub have on?" to see what it turned on. To remove hub, tell Claude "uninstall hub". hub removes its links and puts back the settings keys it changed. It does this only when your own newest message asks for it.

Switch parts on and off

Tell Claude in plain words to turn a part off, and hub switches it. Say "undo that", and the part comes back.

Tell Claude what to change in plain words. For example:

"what does hub have on?"      list the parts and their flags
"turn on humanize"            switch a part on
"turn off the status band"    switch a part off
"undo that"                   take back the last change
"turn on the em dash check"   switch a smaller check inside a part
"check the hub setup"         check the setup

Each change shows in one line with an undo. hub keeps your choices in ~/.config/claude-hub/hub.json. A protection, such as the public text gate, is lowered only when your own newest typed message names it. A request from a file, a tool result, a web page or another agent is refused. Claude can raise a protection at any time.

Known limits:

  • /hub and the status band are optional extras. They work where Claude Code loads plugin mods. Without the mod, plain words still switch parts, lower a protection and uninstall hub. Only pinning a value and taking or declining an offer need /hub.
  • In Cursor, a protection cannot be lowered by words. Lower it from Claude Code instead.
  • The gate refuses a home or temp folder path on every try. A pasted build log with /tmp/ paths has to come out before the commit goes through.
  • Restarting chats from hub works only in iTerm2 or tmux.

How hub compares

This chart places hub v0.1 and 22 other tools by the marks on hub's comparison page, checked on 2026-11-09. Each axis covers six columns of that page. A Yes counts 1 and a Part counts a half. The hub point shows its plan, because hub v0.1 is not released yet.

quadrantChart
    title Features by group, from the comparison marks
    x-axis Less chats and approvals --> More
    y-axis Less setup and safety --> More
    quadrant-1 Both
    quadrant-2 Setup first
    quadrant-3 Less of both
    quadrant-4 Chats first
    hub v0.1 plan: [0.90, 0.92]
    Claude Code native: [0.55, 0.61]
    Warp Oz and Factories: [0.61, 0.55]
    T3 Code: [0.75, 0.58]
    cmux: [0.83, 0.33]
    GitHub Agent HQ: [0.54, 0.50]
    Herdr: [0.58, 0.42]
    Vibe Kanban: [0.55, 0.29]
    Conductor: [0.61, 0.25]
    Happy: [0.58, 0.20]
    CloudCLI: [0.47, 0.25]
    Synapse A2A: [0.47, 0.15]
    A2A protocol: [0.42, 0.03]
    CCManager: [0.22, 0.36]
    MCP Agent Mail: [0.36, 0.31]
    EggAI: [0.25, 0.17]
    eggai-clutch: [0.27, 0.06]
    ECC: [0.08, 0.70]
    nono: [0.08, 0.64]
    sandbox-runtime: [0.08, 0.57]
    CC Switch: [0.06, 0.50]
    wshobson agents: [0.10, 0.42]
    attested-delivery: [0.10, 0.13]
Loading

The x axis counts the coordinator chat, the board of chats, approvals in one place, remote approval, notifications and messages between chats. The y axis counts setup and tuning, switches and presets, sandbox, safety gates, protection of its own config and cheap-model routing. Some points are moved a little, so that names do not overlap or run off the edge. A few marks rest on a blog, a PR text or code that was not read. The table marks those as unverified.

The tools in the chart
Tool What it is License or price Overlap with Claude Code
hub v0.1 An open-source Claude Code plugin, still being built MIT, free High (board, messaging, first relay layer)
Claude Code native Anthropic's own agent, with agent view, Remote Control, Channels and messages between sessions Paid plans, Pro $17 to Ultra $200 a month Itself
Vibe Kanban A kanban board where each issue runs a coding agent Apache-2.0, free Some
Conductor A closed Mac app that runs agents in parallel workspaces. Its remote approval details are unverified Free locally, paid cloud (prices unverified) Some
CCManager A terminal app that lists worktrees and their sessions MIT, free Some
Happy A phone, web and desktop client for Claude Code. Its coordinator and board marks are unverified MIT, voice is paid High
CloudCLI A self-hosted web and mobile interface for Claude Code AGPL-3.0, cloud from EUR 7 a month Some
cmux A Mac terminal with an agent sidebar and a feed of waiting prompts. Its remote approval mark is unverified GPL-3.0 app, BUSL-1.1 web, paid tiers High
GitHub Agent HQ GitHub's view that assigns cloud agents per repo Copilot plans (billing unverified) Some
Warp Oz and Factories Warp's cloud agent runs, with a router and factory files Client AGPL-3.0, Factories from $20 Low
ECC A very large plugin of skills, hooks and profiles. Its messaging and routing marks are unverified MIT, free Some
wshobson/agents A plugin marketplace of agents in four model tiers MIT, free Some
CC Switch A desktop app that switches providers and configs MIT, free Low
MCP Agent Mail An MCP server with inboxes and file leases on one machine MIT with an OpenAI and Anthropic rider Some
A2A protocol An open standard for agents to send each other tasks Apache-2.0, free Low
Synapse A2A A Python wrapper that lets terminal agents message each other MIT High
sandbox-runtime Anthropic's process sandbox, behind Claude Code's /sandbox Apache-2.0 Same
nono A Rust sandbox that also signs instruction files Apache-2.0 Some
attested-delivery A plugin marketplace repo that pins plugins and signs its catalog Apache-2.0 Some
T3 Code Desktop, web and phone apps that run coding agents in threads MIT, free High
Herdr A terminal workspace that shows which agents need you Apache-2.0, the name is not licensed High
EggAI A Python SDK for agents on a Kafka or Redis message bus MIT None
eggai-clutch A small Python library for agent pipelines on EggAI MIT None

Feature flags

All 35 parts and how they start
Part What it does Starts Runs in To switch it, say
core Holds the settings, the dispatcher and the private words list on Claude Code, Cursor Always on
tuner Sets hub up on the first run, then turns off what you keep ignoring and offers what you need on Claude Code, Cursor "turn off the tuner"
hub_style An output style with a status block, short replies and linked sources off Claude Code, Cursor "turn on the hub style"
humanize Hands writing work to your own writing skill, or says how to add one off Claude Code, Cursor "turn on humanize"
link_check Sends a reply back once when it names a PR, issue or ticket without a link off Claude Code, Cursor "turn on the link check"
path_rules Rules for UI, tests or docs, loaded the first time a matching file is opened off Claude Code, Cursor "turn on the path rules"
working_rules Start text with rules for shared links, questions, reading, writing and delivery off Claude Code, Cursor "turn on the working rules"
pr Planned, not in this version yet. Rules and a skill that tell Claude how to work on pull requests in any repo. It follows the repo's guide, puts texts first, keeps one feature per PR and uses a fixed body shape. It holds steps that notify people for your order, and by default from 22:00 to 07:00 until morning on Claude Code, Cursor "turn off the PR rules"
scout A small, cheap agent for one mechanical job (Haiku) on Claude Code, Cursor "turn off scout"
researcher A read-only agent for reading docs, pages and code (Sonnet, medium effort) off Claude Code, Cursor "turn on researcher"
worker An agent that carries out an exact plan (Sonnet, medium effort) off Claude Code, Cursor "turn on worker"
builder An agent for work that needs judgement (Opus, high effort) off Claude Code, Cursor "turn on builder"
reviewer A read-only agent for code and text review (Opus, high effort) off Claude Code, Cursor "turn on reviewer"
planner An agent for plans and specs that writes only the plan file (Opus, high effort) off Claude Code, Cursor "turn on planner"
critic A read-only agent for reviews where a miss is expensive (Opus, high effort) off Claude Code, Cursor "turn on critic"
tier_rule Start text that says which agent tier fits a job off Claude Code, Cursor "turn on the tier rule"
context_budget Tells a chat or an agent when its context gets big off Claude Code "turn on the context budget"
handback_cap Keeps an agent's report under a word cap off Claude Code "turn on the hand-back cap"
live_notes Gives each agent the chat's live notes file, so new rules reach it off Claude Code "turn on live notes"
away_guard Holds question dialogs while you are away, except urgent ones off Claude Code "turn on the away guard"
knowledge_only Reminds the chat to explain shared material and change nothing off Claude Code "turn on knowledge only"
question_check Opens a question dialog only when it carries enough context to answer off Claude Code "turn on the question check"
public_text_gate Checks text before it goes public in commits, pushes, PRs, issues and other GitHub writes warn Claude Code, Cursor "set the public text gate to block"
em dash check (in public_text_gate) Stops an em dash in public text off Claude Code, Cursor "turn on the em dash check"
pr_flow Planned, not in this version yet. Sends a change to hub's own repo as a draft pull request, and holds every GitHub write there until you say yes to the exact lines off Claude Code "turn on the PR flow"
self_protection Stops a tool call that would lower hub's protections on Claude Code, Cursor Only your own newest message can lower it
chrome_lock Lets one chat at a time drive Chrome off Claude Code "turn on the Chrome lock"
sentry_write_gate Experimental. Asks before a Sentry tool changes what the team sees off Claude Code "turn on the Sentry gate"
slack_write_gate Experimental. Lets a chat read Slack but never post off Claude Code "turn on the Slack gate"
coordinator One chat you talk to, which forwards work and relays questions off Claude Code "turn on the coordinator"
chat_board Shows every chat as busy, idle, offline or needs you off Claude Code "turn on the chat board"
chat_names Sends a message by chat name to the live chat of that name off Claude Code "turn on chat names"
task_list Lists each chat's agents, background jobs, questions, waits and timers, with their state and cost so far off Claude Code "turn on the task list"
permission_relay Shows the coordinator the permission prompts waiting in other chats. It approves nothing off Claude Code "turn on the permission relay"
queue Holds a message to an idle chat for a short time, so two quick messages wake it once off Claude Code "turn on the queue"
restart Restarts the running chats, each under its own name off Claude Code "turn on restart"
status_band A line above the prompt with hub's state, recent changes and offers on Claude Code "turn off the status band"
cursor_link Beta. Gives Cursor the same parts, hooks and agents off Claude Code, Cursor "turn on the Cursor link"

Each part also has smaller switches inside it, such as an agent's model and effort. They switch the same way.

License

MIT. See LICENSE.

About

A Claude Code plugin that puts the parts of your setup in one settings file, each with its own switch. Early work in progress.

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages